11/16/2025
A study by researchers from Stanford and Carnegie Mellon has found that AI models are 50% more sycophantic than humans. Not only that, participants rated flattering responses as higher quality and wanted to use them more.
It gets even worse: the flattery made participants less likely to admit they were wrong — even when confronted with evidence. “This suggests that people are drawn to AI that unquestioningly validates, even as that validation risks eroding their judgment and reducing their inclination toward prosocial behavior,” the researchers wrote. “These preferences create perverse incentives both for people to increasingly rely on sycophantic AI models and for AI model training to favor sycophancy.” https://arxiv.org/abs/2510.01395
In other words, humans are hard-wired for approval, and so is AI. It’s a win-win for both sides — a flattery perpetual motion machine.
AI models have turned into high-tech versions of courtesans — mistresses, or prostitutes, found in royal and aristocratic courts in Europe and Asia over the centuries. Among other talents, they often used flattery to seduce and gain status. Now we’re all royals, being sweet-talked by courtesans at the touch of a button.
As Disraeli said, “Everyone likes flattery; and when you come to Royalty you should lay it on with a trowel.”
There are some obvious pitfalls to having flattery applied to our queries with a digital trowel. OpenAI now has 800 million weekly users. And people increasingly trust AI to give them advice on more and more aspects of their lives. Surveys have found that 66% of Americans have used AI for financial advice, that nearly 40% trust AI on medical advice, and that 72% of teens have used AI companions.
Sure, AI will answer any question, but how useful is that when it’s just telling us what we want to hear?
What it’s not doing is what a human advisor or a trusted friend would do: tell us when we’re wrong. Nor, in its eagerness to please, will it tell us when we’re asking the wrong question altogether.
As Plutarch put it, “I do not need a friend who changes when I change and nods when I nod; my shadow does that much better.”
The risk is that instead of being a defense against the echo chambers of social media, AI just becomes a more powerful version.
